This study proposes a novel moving force neural identifier (MFNI), a deep neural network designed for accurately estimating dynamic vehicular axle loads without knowing the vehicle speed and axle configurations (i.e., axle number and spacing). The MFNI architecture comprises a multi-time resolution encoder/decoder (MRED) and a dynamic moving force correction block (DMF-corr block). Firstly, the MRED with hierarchical resolution is developed to adapt to the complex dynamic inverse problem-solving. It can capture the multi-time resolution features of bridge responses and dynamic moving forces, and perform stably under various lengths of identification sampling windows. Then, the DMF-corr block limits the range of identified values by assuming that the vehicular excitation on the bridge is unidirectional, which enables the MFNI model to better determine the axle number, vehicle speed, as well as whether the axle is acting on the bridge. Numerical studies are implemented to confirm the accuracy and efficiency of the proposed MFNI approach, and a parametric sensitivity analysis is conducted to evaluate the robustness of the approach. Finally, a laboratory experiment confirms the practicability of the MFNI in detecting vehicular moving forces with a limited number of sensors and training samples. Both numerical and experimental results demonstrate the potential of the proposed method to serve as a nothing-on-road bridge weigh-in-motion system in practical engineering. •A deep-learning method for moving force identification is proposed, without the prior information on the axle configurations.•The proposed framework can adaptively determine the speed, axle count, and axle base during the moving force identification.•The feasibility of using simulated data for model training and subsequently predicting experimental loads is demonstrated.